Rolling bearing performance degradation trend prediction based on RCMFDE and STAGCN

Abstract

Prediction of rolling bearing performance degradation trends is crucial for ensuring the reliability and safety of rotating machinery in industrial applications. Aiming at deep learning methods such as recurrent neural network, convolutional neural network, and their variants only focus on a single temporal or spatial dimension of time series and lack joint modeling of spatio-temporal features, this paper proposes a rolling bearing performance degradation trend prediction method based on Refined Composite Multiscale Fractional Diversity Entropy (RCMFDE) and Spatio-Temporal Attention Graph Convolution Network (STAGCN). First, the fractional calculus is introduced into the Diversity Entropy (DE), and the multi-scale analysis is extended to propose a feature indicator construction method based on RCMFDE. The spatial correlation of the RCMFDE graph model is parsed by Chebyshev graph convolutional network (ChebGCN), and the gated recurrent unit is designed to learn the temporal correlation. Meanwhile, the weight allocation of key degenerate nodes is dynamically optimized by using the soft-attention mechanism, and the STAGCN prediction framework is established. Experimental verification shows that by improving DE, the robustness of the feature indicator has been enhanced. Compared with the Temporal Fusion Transformer, Efficient Additive Attention Transformer, Series-Core Fused Time Series, TimeXer, and Extended Long Short-Term Memory-Informer models, the MAE and Score of the STAGCN prediction model are reduced by 14.19%–48.13% and increased by 6.42%–14.77%, respectively.